Projects and base models

A project holds everything for one base model: sources, examples, versions, training runs, evaluations, experiments and deployments. Because everything in it uses the same model, its scores always compare like with like. To try the same data on another model, create a second project and import the sources there too.

Every member of an organization sees all its projects.

Before you start

• You are a member or an owner. Viewers can open projects but not change them. See roles.
• Your organization has its storage connected. Until then, New project is unavailable. See Connections.
• Your plan has room for another project. See Plans and billing.

Choose a base model

The base model is the Hugging Face model your adapters are trained for and served with. Enter its ID as Hugging Face shows it, such as Qwen/Qwen2.5-1.5B-Instruct. The model must be:

• Public and not gated. Your Hugging Face token is used for datasets and publishing, never to download base models.
• A chat model, with a chat_template in its tokenizer_config.json. If a model has none, use its instruct or chat version.
• Ideally a supported family, as listed under architectures in its config.json. See Supported models. You can choose a model of any other family: the readiness checks then warn that it may not train or serve.

The model is not looked up when you create the project. The readiness checks before your first experiment, training run or deployment check it and pin it to its current commit, so later changes on Hugging Face don't change your results. A typo only shows up then, so copy the ID from the model's page. Passing the checks does not guarantee the model fits a given GPU. See Supported models.

Create a project

1. Open the project switcher at the top of the sidebar and choose New project.
2. Under Use case, choose the closest match, or Start blank.
3. Enter a Project name and the Hugging Face base model.
4. Adjust the Goal and Generation instructions if you like.
5. Choose Create project.

The project's fields

Use cases

See Recipes and Evaluations.

Change a project's settings

1. On the project's Home, choose Project settings. You can also open Settings at the foot of the sidebar: the project's General settings open first.
2. Change what you need and choose Save settings.

Changing the use case here does not rewrite the goal or instructions. Generation jobs already queued keep the text they started with.

When the base model can no longer change

You can change the base model until the project has an evaluation, a training run, a deployment or an experiment. Sources, examples, recipes and versions don't lock it, so you can prepare data first and settle the model later.

Switch projects

Open the project switcher at the top of the sidebar, type in Find a project… and choose one. You can also press ⌘K (Ctrl+K) and choose Switch to a project.

Delete a project

Owners can delete a project they no longer need. It's deleted at once and can't be undone.

Before you delete

• The owner role. See roles.
• Nothing in the project runs. Wait for or cancel its training runs, experiments, imports and other background jobs, and stop its deployments. A run or deployment whose GPU is still being released, or whose GPU release failed, also counts as running.
• No training run waits for recovery. A run that needs recovery keeps its stopped GPU machine, and its output, until you recover its adapter or cancel it.
• No published models from the project. Stop and delete them on the API screen first. See Publishing.
• No running evaluation in another project uses its deployments. An evaluation of another project that calls one of this project's deployments, as its model or its judge, must finish first.

Steps

1. On the project's Home, choose Project settings.
2. Under Delete project, choose Delete project.
3. Type the project's name exactly as shown.
4. Choose Delete project.

Another of your projects opens. If none is left, the console offers to create one.

What's removed and what stays

Removed:

• the project's sources, examples, recipes, versions, evaluations, experiments, training runs and deployments;
• the files Tensorant wrote for the project on your network volume, such as imported sources, versions and adapters. They are deleted shortly after the project; if the volume cannot be reached then, Tensorant tries again later.

Stays:

• models you pushed to Hugging Face;
• your other projects, connections, API keys and members;
• the dataset imports and training runs the project used this month. They still count toward your plan.

The project's slot in your plan frees at once. The deletion is recorded in the audit trail.

Limits

Only owners can delete a project. See Delete a project.

When something goes wrong